Cities regulate stormwater using registries that record how much of each parcel is paved, roofed, or otherwise impermeable. These registries are updated on slow permitting and planning cycles, while paving happens continuously. The gap matters because small amounts of unrecorded hard surface do real damage. Headwater streams degrade at impervious fractions below 8%, and land-cover change raised peak discharge by 49.5% in one modeled catchment, more than twice the effect of projected climate change (SWMM headwater study). Flow sensors in drains and streams do record the consequences. The difficulty is reading them backward: when a city calibrates its drainage model to observed flow, the fit can stay good even with a wrong land-cover map, because roughness and catchment-width parameters quietly absorb the error (UAV imperviousness study). From a single city's data, "more pavement" and "slightly different pipe and surface parameters" look the same.
We propose to break that ambiguity by learning across many cities without pooling their data. Federated learning trains one shared model by sending model updates, not raw data, between data owners and a coordinator. We use it to train an inference network that reads rainfall and outflow records and returns a posterior over the physical parameters of each drainage sub-area, with the effective impervious fraction as the quantity of interest. Utilities will not share pipe schematics, parcel footprints, or sensor archives. They can, however, contribute to a model that has seen many network geometries and many real storms. That breadth acts as a learned prior on the nuisance parameters, so it can pin down impervious fraction where local calibration cannot. A sub-area whose inferred imperviousness confidently exceeds its registered value is flagged for inspection. Regulators then get evidence on the timescale of storms rather than planning cycles.
Core Claim
Effective impervious fraction cannot be identified from one utility's flow record alone, but it becomes identifiable from a population-level model trained jointly across utilities. Federated training can reach that population model without exposing network geometry or land-use data. The claim can be checked directly. In simulation, local calibration's error on impervious fraction should stay flat as its fit to flow improves. The federated estimator's error should fall as the number and diversity of participating clients grows, and it should approach a centralized model trained on the same data.
The three directions play different roles. Urban impermeability supplies the physics and the target quantity. Federated learning is the mechanism that makes cross-city pooling possible under the data-sharing barriers the file documents. Regulatory lag defines the application: the output is a stale-registry alarm. Our contribution to that literature is a concrete instance of the sensor-to-institution feedback loop that Embedding AI in society calls for but does not build. It is also a way through the evidence dilemma described in the First International AI Safety Report, where regulators must choose between acting on incomplete evidence and waiting for harm.
Relation to Prior Work
The closest work is Enabling Uncertainty-Aware Time-Series Forecasting in Federated Learning for Urban Water Dynamics. It federates a transformer forecaster across utilities that each own disjoint sensors, adds quantile outputs, and reports a 55x reduction in communication and an "Anchor" effect in which one dominant utility stabilizes training. That work forecasts water levels as abstract time series. It has no notion of surface type, no physical parameters, and no output a regulator could act on. We reuse its setting: utilities as clients, sensor-disjoint data, and skewed client sizes. We change the target from future water level to the identification of land-surface parameters, and the model from a black-box forecaster to an inference network tied to a drainage simulator.
Three hydrology papers supply the method and motivate the gap. An intelligent SWMM calibration method narrows calibration search spaces with Latin Hypercube sampling and self-organizing maps. It also shows that which parameters control runoff switches around 10–20% imperviousness, so whether impervious fraction is identifiable should depend on regime. Inundation Modeling and Bottleneck Identification calibrates the Storm Water Management Model (SWMM, the standard US EPA drainage simulator) with a genetic algorithm. It shows that river backwater, not local pipe capacity, often drives flooding. That warns us that a runoff increase need not mean more paving, so our posterior must keep pipe and boundary effects as separate explanations. Both papers are centralized and single-site.
Transfer Learning Using Inaccurate Physics Rule predicts correction coefficients for ungauged basins from static descriptors across CAMELS basins. Our idea is related: a population-level model constrains parameters at sites with too little data. Their setting is centralized, daily, and rural. Ours is federated, sub-hourly, and urban.
The UAV imperviousness study plays two roles. It states the calibration paradox we attack, and its 94–96% accurate image classifier is our strongest non-hydrological competitor, discussed below.
Research Program
Each participating utility holds a drainage network, the sub-catchments draining into it, registered impervious fractions, rain gauges, and flow or level sensors. Each utility runs SWMM locally on its own network. It samples parameter vectors (impervious fraction, width, Manning roughness, depression storage, infiltration parameters, and a pipe-condition factor standing in for siltation) and simulates hydrographs under its own historical rainfall. These simulations never leave the utility.
A shared inference network $q_\phi(\theta \mid r, y, s)$ is trained federatedly with Federated Averaging. Here $\theta$ is the sub-catchment parameter vector, $r$ and $y$ are an event's rainfall and outflow series, $s$ are coarse static descriptors (area, slope, network order, but deliberately not the registered impervious fraction), and $\phi$ are the shared weights. The local objective combines simulation-based posterior learning with a reconstruction term on real storms:
$$ \mathcal{L}_k(\phi) = -\mathbb{E}_{(\theta, r, y) \sim \text{sim}_k} \log q_\phi(\theta \mid r, y, s) \;+\; \lambda\, \mathbb{E}_{(r, y) \sim \text{real}_k}\, \mathbb{E}_{\theta \sim q_\phi}\, \big\| y - \hat{y}_k(\theta, r) \big\|^2 . $$
In this objective, $k$ indexes the utility, $\text{sim}_k$ and $\text{real}_k$ are its simulated and observed events, and $\hat{y}_k$ is a fast learned emulator of that utility's SWMM model. The emulator stays local and is used only so the second term can be differentiated. The weight $\lambda$ trades the sim-to-real correction against simulation fidelity.
The first term teaches the network which hydrograph features separate impervious fraction from width and roughness across many geometries. The second term adapts it to real sensor noise and unmodeled processes. The flag for sub-catchment $j$ is $P_q(\theta_{\text{imp},j} > \theta^{\text{reg}}_{\text{imp},j} + \delta)$, the posterior probability that its effective impervious fraction exceeds the registered value by more than a tolerance $\delta$. That probability is aggregated over recent storms.
Four design questions define the research. The first is identifiability: under what network and rainfall conditions does the federated posterior on impervious fraction narrow while a local one does not? The 10–20% sensitivity switch predicts this will differ between lightly and heavily paved areas. The second is client composition: does the Anchor effect help or hurt here, given that a dominant utility may cover a narrow range of geometries? We compare data-weighted averaging with diversity-weighted and clustered aggregation. The third is attribution: does the pipe-condition parameter correctly absorb siltation and backwater cases, so they are not flagged as paving? The fourth is adversarial participation. A utility with undeclared paving has a reason to send manipulated updates that shrink its own flags. We run the method under secure aggregation and test poisoning with the client-tracing tools in FLForensics. This addresses the trust gap that the AquaCast paper explicitly leaves open.
Baselines
The naive reference is to trust the registry. Its detection rate for undeclared change is zero by construction, and its imperviousness error equals the injected change. This baseline fixes the scale of the problem.
The strongest physics-based competitor is local calibration. Each utility calibrates SWMM on its own events with the self-organizing-map and Latin Hypercube procedure of An intelligent SWMM calibration method, with impervious fraction left free, and compares the result to the registry. That paper reports verification NSE of 0.777 and an uncertainty R-factor reduced from 0.396 to 0.226 on its Changchun campus. NSE (Nash–Sutcliffe efficiency) is the standard hydrograph-fit score, where 1 is perfect. The genetic-algorithm calibration of Inundation Modeling and Bottleneck Identification reports NSE of 0.876 and 0.826. These are flow-fit numbers. For the error in recovered impervious fraction, or for detecting undeclared change, no reported result was found on Lacuna for any method. We will therefore produce that baseline number ourselves. Local calibration gets the same events, the same simulator, and a simulation budget equal to each client's share of the federated budget. Beating it means lower impervious-fraction error at matched or better flow NSE. Matched NSE matters here: the calibration paradox predicts that local calibration fits flow well while getting imperviousness wrong.
The strongest machine-learning competitor is a federated forecaster with residual alarms. We train the AquaCast-style federated quantile forecaster from Enabling Uncertainty-Aware Time-Series Forecasting on pre-change data. A change is flagged when observed flow persistently exceeds its upper quantile. That paper reports a Scaled Quantile Loss of 0.1743 against 0.2669 for a median baseline on Lausanne; we will reproduce this to verify our implementation. This baseline can detect that something changed, but it cannot say how much or whether the cause is paving. We compare detection AUC and false flags on pipe-degradation events at equal federated rounds and parameter count.
The upper bound is a centralized version of our inference network with all simulations and events pooled. The federated version should come within a small margin of it.
Two ablations isolate the key ingredient. The first trains the same inference network locally with no federation, which tests whether the population prior is what breaks equifinality. The second removes the pipe-condition and backwater terms, which tests whether attribution depends on modeling other explanations explicitly.
Why not just…?
Why not just map impervious surfaces from imagery? This is the most serious alternative. The RQE-plus-boosting classifier in the UAV imperviousness study reaches 94–96% land-cover accuracy, and satellite imagery is cheap. Imagery measures what surface is visible, not how much of it drains to the network. A roof draining onto a lawn counts as paved in imagery but not in runoff. Imagery also says nothing about pipe condition, and the file notes that private-property flyovers face legal limits. We do not argue against imagery; we measure the overlap. On Wartegg, where UAV maps and 36 storm events coexist, we compare imagery-only estimates, runoff-only estimates, and a fused estimator that uses imagery as a prior. If runoff inference adds nothing beyond imagery for detecting hydrologically relevant change, we will say so and recast the method as a check on imagery-derived registries rather than a replacement.
Why not just calibrate each city's model better? The calibration paradox says better calibration of flow does not imply better recovery of imperviousness. The local-calibration baseline tests this directly. The prediction is that its imperviousness error stays roughly constant as its NSE rises, while the federated estimator's error falls with the number of clients. If local calibration with a tight, hand-set prior on width and roughness matches us, the value lies in the prior, not in federation. To find out, we include a baseline that uses a literature-informed fixed prior.
Why not just centralize the data? Where centralization is legally possible, it is the upper bound and should be used. The file documents why it often is not: utilities withhold schematics and sensor archives, and governance rules diverge across jurisdictions (Navigating Governance Paradigms). Our question is how much of the centralized benefit survives federation. The federated-versus-centralized gap answers it.
Proposed Experiments
The decisive experiment is synthetic, because only there is the true impervious fraction known after a change. We build a population of 50–100 SWMM networks with varied geometry, slope, and paving from 0% to 90%, driven by Swiss rainfall in the spirit of the MeteoSwiss-synthesized networks used by the AquaCast paper. We group them into utilities with Zipf-skewed sizes. Into held-out sub-catchments we inject undeclared impervious increases of 1–15 percentage points, varying their position (headwater, midstream, downstream) because the SWMM headwater study found midstream paving the most damaging. Separately we inject pipe-siltation and downstream backwater events as confounders. We report detection AUC, the smallest change detected at a fixed false-flag rate, impervious-fraction error, and posterior calibration. Each is plotted against number of clients and paving regime. The core claim fails if the federated estimator's error does not fall below local calibration's, or does not shrink as clients are added.
Real-data checks follow, with honest limits. The five Iowa headwater watersheds span 5.3% to 37.1% impervious area with 5-minute flow records. We treat each as a client, hold one out, and test whether runoff-inferred imperviousness ranks and approximates the land-use-derived values. Five sites support a sanity check, not a strong statistic. Wartegg (36 events, UAV and professional orthophotos) tests recovery against an independent map and supports the imagery comparison. The Changchun campus (26 sub-catchments, 0–94% paving) tests whether the identifiability boundary sits near the reported 10–20% switch. The Mugu basin tests whether backwater cases are correctly not attributed to paving. Real data for an observed change over time is missing from the file. A retrospective pilot with a utility that has dated paving permits would be the first deployment step.
The robustness check repeats the synthetic experiment with one utility sending updates crafted to suppress its own flags, both with and without secure aggregation and client tracing. Success means the manipulation is either detected or bounded in effect. Failure would mean the method cannot be used where liability is at stake, which the file suggests is exactly where it would be deployed.
References
- Using the Storm Water Management Model to predict urban headwater stream hydrological response to climate and land cover change
- High-quality observation of surface imperviousness for urban runoff modelling using UAV imagery
- Embedding artificial intelligence in society: looking beyond the EU AI master plan using the culture cycle
- The First International AI Safety Report
- Enabling Uncertainty-Aware Time-Series Forecasting in Federated Learning for Urban Water Dynamics
- An intelligent SWMM calibration method and identification of urban runoff generation patterns
- Inundation Modeling and Bottleneck Identification of Pipe–River Systems in a Highly Urbanized Area
- Transfer Learning Using Inaccurate Physics Rule for Streamflow Prediction
- Navigating Governance Paradigms: A Cross-Regional Comparative Study of Generative AI Governance Processes & Principles